Cross-checks in the open
Open isn't enough — an encoding you can read but can't test is still a claim. We compare our rules with other calculators and datasets, and publish the comparisons and the code that runs them. Max Ghenis is CEO of both Axiom and PolicyEngine, the oracle we compare against most, and PSL Foundation fiscally sponsors both. In September 2026, our coverage map tied 14,030 of our 34,810 US rules to a program that a live comparison exercises, though a comparison of a program does not check each of its rules; a public file gives each rule's status.
The oracle dashboard
Open full sizeCross-engine comparisons as they stand, including divergences we have not yet explained.
The oracles
PolicyEngine
Open-source microsimulation of US and UK tax and benefit policy, co-founded by Max Ghenis.
Visit the oracleTAXSIM
NBER's tax calculator, the reference standard in economics research for US income tax liabilities. PolicyEngine is building TAXSIM's successor with its author's cooperation, and our runs use the TAXSIM executable that PolicyEngine packages.
Visit the oracleUKMOD / EUROMOD
The EUROMOD family of tax-benefit microsimulation models, maintained by university and EC research teams.
Visit the oracleSPSD/M
Statistics Canada's Social Policy Simulation Database and Model — the reference microsimulation of Canadian federal and provincial taxes and transfers.
Visit the oracleSOUTHMOD
UNU-WIDER's tax-benefit microsimulation models for countries in the Global South, run on the EUROMOD software. 40 suites compare our rules for Ghana, Uganda, Zambia, Ethiopia, and Rwanda: 242 of 245 comparisons match, and we attribute the other three to gaps in the models, with the arithmetic published. UNU-WIDER's licence bars giving the models to anyone else, so no CI runner can hold them: we run these suites by hand on our licensed machine. Every household is synthetic, and the SOUTHMOD bundle has no Rwandan microdata at all. None of the Axiom rules these suites compare went through our encoder's apply step: each was attested by hand or carries no encoding record.
Visit the oracleSNAP quality-control data
USDA's case-level QC microdata. For 5,175 reviewed FY 2024 households in six states, the research firm Mathematica computes a benefit for USDA from the edited case record; we compute it from the same record, taking income and several deductions as given, and compare. The file keeps only eligible households, so the replay checks benefit arithmetic and leaves eligibility untested.
Visit the oracleTwo more are not connected yet: PSL Tax-Calculator and ACCESS NYC have no published comparison with our encodings.
SOUTHMOD acknowledgement: The results presented here are based on the tax-benefit microsimulation models for Ghana (GHAMOD), Uganda (UGAMOD), Zambia (MicroZAMOD), Ethiopia (ETMOD), Rwanda (RWAMOD) in SOUTHMOD_A4.0. Models in the SOUTHMOD bundle are developed, maintained and managed by UNU-WIDER in collaboration with SASPRI (Southern African Social Policy Research Insights), the International Inequalities Institute at the London School of Economics and Political Science, and local partners in selected developing countries (Bolivia, Colombia, Ecuador, Egypt, Ethiopia, Ghana, Mozambique, Peru, Rwanda, Mainland Tanzania, Uganda, Viet Nam, Zambia, and Zanzibar) in the scope of the SOUTHMOD project. The results presented here are based on EUROMOD version EM_Executable 1.0.0 (run through the euromod Python connector 0.2.18). Originally maintained, developed and managed by the Institute for Social and Economic Research (ISER), since 2021 EUROMOD is maintained, developed and managed by the Joint Research Centre (JRC) of the European Commission, in collaboration with EUROSTAT and national teams from the EU countries. We are indebted to the many people who have contributed to the development of SOUTHMOD and EUROMOD. The results and their interpretation presented in this publication are solely the Axiom Foundation's responsibility. Input data: the policy systems run under the input-dataset configurations UNU-WIDER built from these surveys; only each dataset's variable list is read, and every comparison household is synthetic. GHAMOD: Ghana Statistical Service (GSS) (2020). Ghana Living Standards Survey Round 7 (GLSS7), 2016–2017. Accra: GSS. https://microdata.fao.org/index.php/catalog/1397 UGAMOD: Uganda Bureau of Statistics (UBOS) (2025). Uganda National Household Survey, 2023/24 (UNHS 2023/24). Kampala: UBOS. https://microdata.ubos.org:7070/index.php/catalog/80/download/275 MicroZAMOD: Zambia Statistics Agency (ZamStats) (2024). Living Conditions Monitoring Survey, 2022 (LCMS 2022). Lusaka: ZamStats. https://www.zamstats.gov.zm/wp-content/uploads/2024/07/2022-LCMS-Report-2022.pdf ETMOD: Central Statistical Agency of Ethiopia (CSA) (2024). Ethiopia Socioeconomic Panel Survey, Wave 5 (ESPS-5) 2021–2022. Addis Ababa: CSA. https://microdata.worldbank.org/index.php/catalog/6161 RWAMOD: no input data (the bundle ships none; runs use a header-only file and synthetic rows)
How a comparison runs
Same case, side by side
Concept-keyed test cases — a household, an income, a date — run through Axiom and one or two oracles behind adapters that translate each engine's inputs and outputs.
Normalized outputs
Engine results map onto shared program concepts before comparing.
Disagreements classified
Reports sort mismatches by cause — an encoding gap, an oracle gap, a harness artifact, or an explained residual — and count the ones still unexplained.
The comparison harness is open source. If you maintain a calculator, a screener, or an eligibility system, plug it in and put us to the test.
Looking for the engineering contracts behind this? Documentation →